Unsupervised online anomaly detection in road bridges from single-point acceleration data

Ensuring the safety of civil infrastructure, particularly bridges, is essential. Yet, conventional multi-sensor monitoring systems are often cost-prohibitive for large bridge networks. To address this, we propose an unsupervised online anomaly detection framework that relies on single-accelerometer data. Dynamic characteristics are extracted through random decrement and autoregressive modeling, while Mahalanobis Distance quantifies structural changes with reduced outlier influence. Temperature effects are corrected using a machine learning model. The framework was validated on four bridge types over more than one year of monitoring. Results show that the method effectively distinguishes transient events, such as sensor malfunctions or abnormal amplitudes, from permanent structural changes, exemplified by pavement replacement. In the latter case, the Mahalanobis Distance mean increased from 5.2 to 11.5 (121 %), demonstrating sensitive detection of system changes. These findings highlight the framework's potential as a cost-effective tool for early anomaly warning and preventive maintenance in large bridge inventories. • An unsupervised online anomaly detection framework in bridges is developed. • The detection frame work is based on single-accelerometer data. • Structural changes are quantified using a robust Mahalanobis Distance metric. • An XGBoost model is employed to correct temperature-induced variations.